[experimental/python] Improvements, gaps and fixes: Part#1

[khalatepradnya]

Gaps in features and enhancement:

Must have:

• Build Python wheel on arm64 platform
• Enable remote simulator target (and corresponding tests)
• Enhance tests for remote backend with new Python APIs
• Port vqe module
• Port parallel module
• Port chemistry module
• Fix State, get_state_async, AsyncStateResult
• Enable OptimizationResult
• Enable broadcast in mpi
• Implement observe_async in eager mode
• Bug in AST compilation where register_name='name' is ignored
• bug with ordering of kernel definitions and kernel composition
• bug with transitive kernel dependencies
• sample_async eager mode
• check control self.popValue() calls (bug with too many calls to popValue fixed for one scenario, could be others)
• vqe for eager / mlir mode kernels
• drop qlist
• Implement from_state: Will be addressed separately
• range(q, p, -1), p == 0
• Add Python binding for NvcfSimulatorQPU similar to PyRemoteSimulatorQPU

Nice to have:

• Enable building and usage of old Python bindings and new MLIR bindings side-by-side (fix for LLVM static initialization issue)

[khalatepradnya]

Issues related with State will be addressed by #1123

[khalatepradnya]

• parallel fixed in commit d5459b2 on PR Unify python stack #1077
• vqe fixed in commit 4bf7a4a on PR Unify python stack #1077

[khalatepradnya]

• chemistry fixed in commit c635d9f on PR Unify python stack #1077

[amccaskey]

note here - need to disallow empty list initializer l = [] since we don't allow default empty std::vector<T> construction (no dynamic memory allocation). Might need a new type with size argument for list initialization (beyond constant initialization l = [1,2,3...], something like l = cudaq.list_init(10)

[bmhowe23]

I think the changes to JITExecutionCache.cpp (as originally done in a610481 but later backed out) are likely still needed for thread safety reasons. Is that being tracked somewhere?

[khalatepradnya]

I think the changes to JITExecutionCache.cpp (as originally done in a610481 but later backed out) are likely still needed for thread safety reasons. Is that being tracked somewhere?

Yes, I have it in a new PR. But, good to note here as well for tracking.

[amccaskey]

Note - Need to fix issue with kernel composition requiring kernel argument being defined before kernel taking kernel as input.

[amccaskey]

Fix bug in AST mode where register_name is ignored for measurements

[khalatepradnya]

Remote simulator target tests enabled - PR #1187

[khalatepradnya]

arm64 Python wheel fix - PR #1216

[schweitzpgi]

note here - need to disallow empty list initializer l = [] since we don't allow default empty std::vector construction (no dynamic memory allocation). Might need a new type with size argument for list initialization (beyond constant initialization l = [1,2,3...], something like l = cudaq.list_init(10)

Pedantically speaking, we could allow it. But it wouldn't be terribly interesting. Fast though. Processing a vector with 0 elements should zoom-zoom. (I recall adding code to the bridge to handle 0-element initializer lists, but don't recall exactly the status.)

[1tnguyen]

• Add Python binding for NvcfSimulatorQPU similar to PyRemoteSimulatorQPU

In main, we've added NvcfSimulatorQPU as a derived class of RemoteSimulatorQPU. This inheritance needs to be mirrored in the new Python bindings.

[khalatepradnya]

Error handling and messaging captured in #1251

[khalatepradnya]

#1300